There were eighteen healthy, right-handed participants (6 males, mean age 23.17 years, SD 3.31, range 19–30). All had normal or corrected-to-normal vision, and gave informed written consent to participation in accordance with the University College London Research Ethics Committee.

Participants made two visits to the laboratory, 7 days apart. On their first visit, they performed a 10-min fMRI task completely unconnected to this experiment, before undergoing high resolution structural MRI scanning for an hour. (No functional scanning for this experiment was conducted on their first visit.) After structural scanning they had a break, and then sat at a desktop computer to undertake learning for the experiment presented here. Stimuli were photographs of scenes and photographs of single objects presented on a plain white background. Objects were used because they were matched for the task requirements but, unlike scenes, imagining objects has not been found to be hippocampal dependent (e.g., Hassabis, Kumaran, & Maguire, 2007, Hassabis, Kumaran, Vann, et al., 2007, Hassabis and Maguire, 2007). These pictures were shown one at a time, with a two word caption underneath which described the stimulus (Fig. 1A). Participants were instructed to remember both the picture and the caption for a memory test the following week. Each stimulus was shown for 4 sec. The stimuli were shown five times each with the order of stimuli uniquely randomized for each participant. We based the number of repetitions on pilot data, which suggested that five repetitions was sufficient for reliable subsequent recall one week later. To ensure attention during encoding, each time a stimulus was shown 1 of 5 questions was asked: “Caption match the picture?”, “Indoor or outdoor?”, “Mainly man-made or natural”, “Find this scene/object in this country?” and “Does the picture contain anything red?”. All questions were asked for each stimulus over the course of the five learning trials, and were displayed underneath the stimulus after it had appeared alone on screen for 4 sec. The picture remained on screen while the question was presented and participants then had a maximum of 5 sec to respond using a button press. There were 20 scene stimuli and 20 object stimuli, plus 2 stimuli of each category which had incorrect captions – these were used as lures for the “Caption match the picture?” question. The lures and associated photographs were not used in subsequent scanning.

On their second visit one week later, 30 min prior to fMRI scanning participants learnt a new set of 20 scene and 20 object photographs (plus 2 lures), using the same procedure as their first visit. They were then trained and given practice on the scanning task (Fig. 1B), which proceeded as follows. There were 3 scene conditions (imagine scenes, recall 1-week-old scenes, recall 30-min-old scenes), 3 matched object conditions using single isolated objects instead of scenes, and a fixation baseline condition. The order of trials was pseudo-randomised across participants. In the imagine scene/object conditions, participants were given a two-word cue (in green lettering) describing the scene or object they had to construct in their imagination. They then saw the cue “close eyes”, and had 8 sec in which to construct the scene/object. They had been trained to keep their imagined viewpoint fixed, and to attempt to come up with something new rather than previous memories. An auditory beep alerted them to open their eyes, and two self-paced ratings (5 sec max) were taken. “How vivid?” [1–5, max 5] asked how vivid was the scene/object they had constructed. “Like memory?” [1–5, max 5] asked how similar it was to a specific memory.

In the recall 1-week and recall 30-min conditions, the two word cue displayed at the beginning of the trial (with yellow lettering) had been previously seen as a caption under a photograph during learning. Participants spent the 8 sec after the “close eyes” cue bringing back into their mind's eye the relevant scene or object. Their training emphasised they should not change the stimulus in any way or use their imagination, but just to spend the time focussing on bringing back what they remembered. As before, a rating for vividness was then taken, and participants responded to the “Like memory?” question with how accurately they felt their memory reproduced the original stimulus. If participants could not remember the stimulus at all, they had been trained not to imagine something new, but rather select 1 out of 5 for vividness and 1 out of 5 for “Like memory?”. These trials were then excluded from further analysis.

In the fixation baseline condition, participants were given the cue “white cross”, and had been trained to imagine a small white fixation cross and focus on it for the 8 sec. They then gave a vividness rating, and the “Like memory?” question was replaced with “How focussed?”. Participants responded with the extent to which they had maintained focus during the trial. This baseline was designed to go some way to controlling for the task demands, such as reading and understanding the cue, and maintaining attention.

Immediately on being removed from the scanner, participants were given surprise memory tests to assess whether they could remember the photograph stimuli and the two-word scene construction cues. In the test for the photographic stimuli, they were shown all 80 scene and object stimuli learnt prior to scanning, plus 40 lures they had not seen before (20 new scenes and 20 new objects), and were asked to indicate with a button press “Remember” or “Do not remember” followed by a confidence rating (1–5, max 5). Timing was self-paced with a maximum of 5 sec per stimulus. An identical memory test was then administered for the two-word cues they had seen for scene/object construction during scanning. Participants were shown all 40 two-word cues that had been used during the imagine conditions during scanning, plus 20 lures (10 scenes and 10 objects) which they had not seen before. Finally, participants answered a series of questions about the cognitive strategies they used during scanning – these are detailed in the behavioural results.

Structural and functional data were acquired using a 3T Siemens Trio MRI scanner (Erlangen, Germany). Both types of scan were performed within a partial volume centred on the medial temporal lobe, which enabled the images to be of sufficiently high resolution to delineate the hippocampal subfields (functional scans had a 1.5 mm³ isotropic voxel size and structural scans had a 0.5 mm³ isotropic voxel size).

Structural images were collected using a single-slab 3D T2-weighted turbo spin echo sequence with variable flip angles (SPACE) (Mugler et al., 2000) in combination with parallel imaging, to simultaneously achieve a high image resolution of ∼500 μm, high sampling efficiency and short scan time while maintaining a sufficient signal-to-noise ratio (SNR). After excitation of a single axial slab the image was read out with the following parameters: resolution = .52 × 0.52 × .5 mm³, matrix = 384 × 328, partitions = 104, partition thickness = .5 mm, partition oversampling = 15.4%, field of view = 200 × 171 mm², TE = 353 msec, TR = 3200 msec, GRAPPA ×2 in phase-encoding (PE) direction, bandwidth = 434 Hz/pixel, echo spacing = 4.98 msec, turbo factor in PE direction = 177, echo train duration = 881, averages = 1.9. For reduction of signal bias due to, for example, spatial variation in coil sensitivity profiles, the images were normalized using a prescan, and a weak intensity filter was applied as implemented by the scanner's manufacturer. To improve the SNR of the anatomical image, four scans (taking ∼12 min each) were acquired for each participant, coregistered and averaged. Additionally, a whole brain 3D FLASH structural scan was acquired with a resolution of 1 × 1 × 1 mm.

Functional data were acquired using a 3D echo planar imaging (EPI) sequence which has been demonstrated to yield improved BOLD sensitivity compared to 2D EPI acquisitions (Lutti, Thomas, Hutton, & Weiskopf, 2013). Image resolution was 1.5 mm³ isotropic and the field-of-view was 192 mm in-plane. Forty slices were acquired with 20% oversampling to avoid wrap-around artefacts due to imperfect slab excitation profile. The echo time (TE) was 37.30 msec and the volume repetition time (TR) was 3.65 sec. Parallel imaging with GRAPPA image reconstruction (Griswold et al., 2002) acceleration factor 2 along the PE direction was used to minimize image distortions and yield optimal BOLD sensitivity. The dummy volumes necessary to reach steady state and the GRAPPA reconstruction kernel were acquired prior to the acquisition of the image data as described in Lutti et al. (2013). Correction of the distortions in the EPI images was implemented using B0-field maps obtained from double-echo FLASH acquisitions (matrix size 64 × 64; 64 slices; spatial resolution 3 mm³; short TE = 10 msec; long TE = 12.46 msec; TR = 1020 msec) and processed using the FieldMap toolbox available in SPM (Hutton et al., 2002).

Data from the post-scan memory test and interview were analysed using repeated measures ANOVAs and paired sample t-tests (SPSS 17.0, Chicago: SPSS Inc.) with a significance threshold of p < .05.

FMRI data were analysed using SPM12 (www.fil.ion.ucl.ac.uk/spm). All images were first bias corrected to compensate for image inhomogeneity associated with the 32 channel head coil (Van Leemput, Maes, Vandermeulen, & Suetens, 1999). Fieldmaps had been collected and were used to generate voxel displacement maps. EPIs for each session were then realigned to the first image and unwarped using the voxel displacement maps calculated above. The four high-resolution structural images were averaged to reduce noise, and co-registered to the whole brain structural scan. EPIs were also co-registered to the whole brain structural scan.

We manually segmented the subfields of the hippocampi using the high-resolution structural image (resolution ∼0.5 mm³) for each subject, following the protocol of Bonnici, Chadwick, Kumaran, et al. (2012) with reference to Duvernoy (1988). We modified the protocol for anterior hippocampus as follows (Fig. 2). Segmentation of the uncus began in the first coronal slices where DG was observed and covered the band of Giacomini (Fig. 2A). The anterior-most section of the uncus, which is associated with the amygdala, was not included. At this level DG was clearly visible, and dorsal to this a hypo-intensity in the T2 signal and narrowing of the gyrus was identified as the transition from CA3 to CA1. More posteriorly the uncus was seen to separate from intraventricular hippocampus, sitting alongside it (Fig. 2B and C). The most posterior slice to include the uncus before it disappeared from view (Fig. 2D) marked the posterior boundary of anterior hippocampus. In these slices, the CA3/CA1 border was placed at the ‘shoulder’ of the gyrus where it began to narrow towards CA1. The definition of the subiculum in anterior (a) hippocampus was adjusted as follows. The aCA1/aSub border was positioned ventral to aDG guided by the study of Ding (2013). A marked hypo-intensity in the T2 signal marked the border between aSub and PrS/PaS. Following the analysis by Fischl et al. (2009), PrS/PaS went no further ventrally than the ‘shoulder’ of the gyrus in order avoid inclusion of entorhinal cortex. Images were segmented using ITK-SNAP (Yushkevich et al., 2006) and a graphics tablet. Segmented masks were then resliced to 1.5 mm isotropic voxels to match the functional images using SPM (with 4th degree B-Spline interpolation). Masks for DG, CA3, CA1 and Sub were then divided into anterior (a) and posterior (p) portions as described above.

For the regions of interest (ROI) analyses, functional data were analysed at the single subject level without warping or smoothing. We built a general linear model (GLM) for each subject with 7 task regressors (imagine scenes, recall 1-week scenes, recall 30-min scenes, imagine objects, recall 1-week objects, recall 30-min objects, fixation task). Each condition was modelled from the onset of the cue to just before the beep sounded indicating the eyes should be opened (10 sec). Additional regressors modelled the beep (as a zero length event) and the ratings period. The ITI was not modelled and so acted as the implicit baseline. Two further regressors modelled the BOLD signal obtained from ROIs in the white matter and CSF. Following estimation of the GLMs, each subject's mask image was inspected for dropout in the hippocampus. In some subjects, voxels were excluded by SPM, which was traced back to single volumes with large spikes. These were repaired by averaging the volumes on either side and modelling any repaired volumes using an extra regressor in that subject's design matrix. To perform the ROI analyses we extracted the mean value from the contrast image for each condition (relative to the implicit baseline). We used t-tests to evaluate significance at the group level. All results presented here are for scene conditions with object baselines subtracted. Parameter estimates for each condition individually versus the fixation baseline task are provided in the supplementary material (Fig. S1).

For group level analyses we created a group DARTEL template (Ashburner, 2007) using the whole brain structural scans from each participant's first visit. Functional images were warped via the group space to MNI space, then spatially smoothed using a Gaussian smoothing kernel of 4 × 4 × 4 mm full-width at half maximum, which we found to be the minimum to satisfy the smoothness assumptions of random field theory-based multiple comparisons correction in SPM. GLMs were created as above (with one modification – the fixation baseline condition was left unmodelled). After estimating the GLMs we summarised results at the group level using one sample t-tests in SPM. Where conjunctions are reported, these are tests against the conjunction null, that is to say a logical AND across conditions. Results are reported using topological FDR correction at p < .05 with cluster forming threshold of p < .01 unless stated otherwise.

Many studies using scene or event stimuli, in the context of various experimental paradigms, have found a significant response in the medial portion of anterior hippocampus (e.g., Addis et al., 2007 [Fig. 1 their paper], Benoit & Schacter, 2015 [Fig 2, their paper], Hassabis, Kumaran, & Maguire, 2007 [Fig. 2, their paper], Zeidman et al., 2015 [Fig. 3 their paper]). Our first aim was to identify the precise hippocampal structures underlying this functional region. We found PrS/PaS to be the only structures engaged both by imagining novel scenes and recalling scenes from the past. What function might PrS/PaS serve?

In rats, dorsal (posterior) PrS and PaS contain a range of cell types that code for space. This includes place cells, which represent an animal's heading-invariant location (Taube, 1995), head-direction cells which represent place-invariant heading (Cacucci, Lever, Wills, Burgess, & O'Keefe, 2004), as well as conjunctive place-by-direction cells. More recently, PrS/PaS have been found to contain grid cells that fire at regular intervals over the environment (Boccara et al., 2010). This is in addition to conjunctive grid-by-heading and border cells, which fire when an animal is close to an environmental boundary. Far less attention has been given to ventral (anterior) PrS/PaS, which is more challenging to record in animals, however the extant findings suggest PrS/PaS could have the capacity to represent spatial properties of imagined scenes in humans.

In humans, several recent studies have capitalised on high resolution functional neuroimaging to investigate the hippocampus and connected structures. Maass et al. (2014) identified a peak of activity for visually perceiving novel scenes in the vicinity of PrS, which may have been representing the scenes while they were being perceived (Zeidman et al., 2015). Using functional connectivity analyses, Maass, Berron, Libby, Ranganath, and Düzel (2015) found distal subiculum, which borders on PrS, to have preferential connectivity with PHC – a region closely associated with scene processing (see also Libby, Ekstrom, Ragland, & Ranganath, 2012). They also found that posterior-medial EC, bordering on PaS, showed a similar profile. Although they did not separately segment PrS or PaS, their results suggest this region may have a preferential response to scenes due to its functional connectivity with PHC.

Understanding the role of PrS/PaS in humans may be helped by considering neuropsychological studies in patients with hippocampal lesions. Hassabis, Kumaran, Vann, et al. (2007) asked patients with bilateral hippocampal lesions and amnesia to describe fictitious atemporal scenes, such as “a sandy beach”. Evaluating patients' responses on a number of indices, they showed a specific deficit in spatial coherence; imagined scenes lacked richness because they were spatially fragmented. Similarly, Mullally et al. (2012) showed patients with hippocampal lesions a scene photograph and asked them to report what they would see if they stepped back from the depicted viewpoint in their imagination. While the patients' responses contained appropriate semantic content, their spatial detail was markedly reduced compared to controls, as was the vividness they reported of the imagined scene. These studies suggest that one contribution of the hippocampus to scenes may be spatial in nature. We propose that the tasks employed in this study – imagining novel scenes and recalling scenes from the past – engage the hippocampus in order to construct internal models or representations of scenes (Hassabis and Maguire, 2007, Maguire and Mullally, 2013). At the core of this proposal is a spatially coherent representation in the hippocampus which binds the scene's elements from across cortex, and our results suggest that anterior PrS/PaS may be key to the spatial aspect of this process.

Having investigated the overall response to scenes in the hippocampus, we next identified the similarities and differences between imagining novel scenes and recalling scenes from the past.

Our experiment included three experimental tasks – subjects imagined novel scenes, recalled scenes they had viewed one week before scanning and recalled scenes they had viewed 30 min before scanning. We hypothesised a similar response in the hippocampus for imagining novel scenes and recalling scenes from a week earlier, as we expected both to place similar demands on the process of scene construction. Although anterior hippocampus was engaged by both conditions as predicted, there were differences in the subregions engaged by each. Imagining novel scenes relative to single objects engaged aSub and PrS/PaS, whereas recalling scenes from a week earlier engaged the uncus and PrS/PaS. Unpacking these results, we first ask why was there a particular response in aSub to imagining scenes?

There were two distinct regions included in our mask for aSub – proximal or prosubiculum (ProS) and more distal subiculum proper, which lies further from DG. The hippocampal subfields project from DG to ProS then subiculum, which in turn projects to cortical and subcortical regions, and for this reason subiculum is considered the main output of the hippocampal subfields (O'Mara, 2006). It may operate to transform spatial codes in CA1 into compressed, information-rich codes suitable for transmission to other brain regions (Kim, Ganguli, & Frank, 2012). In a thorough review of subicular anatomy, Ding (2013) highlighted several features of the ProS and subiculum's connectivity which, we suggest, make them ideal candidates for involvement in constructing scenes. ProS has reciprocal connectivity with perirhinal cortex, which could convey objects and object novelty for inclusion in a scene (Murray & Richmond, 2001). Furthermore it projects to vmPFC; in this study we found vmPFC activation for constructing scenes as well as recalling scenes from a week earlier. It is possible that the enhanced response in vmPFC and other cortical regions was due to items retrieved from separate brain regions being assembled into a coherent scene representation in the hippocampus (Bonnici, Chadwick, Lutti, et al., 2012, Nieuwenhuis and Takashima, 2011). What vmPFC may contribute to memory and imagination is an open question, and there is not yet a consensus on this matter (Maguire, 2014). Also relevant to the connectivity of aSub are direct projections (in monkeys and rodents) from distal subiculum to the RSC and the mammillary bodies (Ding, 2013). These regions are involved in the head direction system, and together may represent and adjust the imagined heading in constructed scenes. Together, the available information on anatomical connectivity of subiculum suggests that transmission of spatial and object information may be of key importance to its function.

Next, we considered our finding that the 1-week recall condition engaged the uncus; indeed this was the only hippocampal subregion with greater activity for 1-week scene recall than imagining novel scenes. The uncus is a complex structure containing modified versions of the same subfields found in the main (intraventricular) hippocampus. Amaral, Insausti, and Cowan (1984) demonstrated that in monkeys, the DG of the uncus has direct commissural connections with the contralateral hippocampus, and these connections terminate specifically in the contralateral uncus. The uncus may therefore support inter-hemispheric connectivity. The subfields of the uncal hippocampus also have anatomical connections with other regions aligned with scene and object processing, including projections from uncal subiculum to the mammillary nucleus (Rosene & Van Hoesen, 1987) and uncal CA1 to perirhinal cortex and pre-frontal cortex (Insausti & Munoz, 2001). We did not have sufficient spatial resolution to distinguish individual subfields of the uncus, and further work is clearly needed to understand the functional implications of the modified subfields it contains.

The 30-min recall condition was associated with significantly less activation of aSub and uncus than the other conditions. One explanation is that this task was simply easier than the others, as reflected by participants' difficulty ratings. However, there is evidence that an interpretation based on difficulty is insufficient to explain our imaging results. Participants' difficulty ratings for the scene tasks were matched in the object baseline conditions, controlling for basic attentional effects. Furthermore, we found robust activation of PrS/PaS and pCA3 for the 30-min recall condition, demonstrating that 30-min recall did engage the hippocampus. In the wider brain there was also significant activation of RSC and PHC for 30-min recall, which we return to shortly. It is our proposal that the need to construct representations of scenes, either novel or recalled from a week earlier, taxed aSub and the uncus and contributed to the participants' sense of difficulty in those conditions. By contrast, in the 30-min recall condition, representations of the scenes had recently been constructed, leading to reduced activation of aSub and the uncus. All three conditions shared the requirement to represent the scenes such that they could be vividly experienced, which was supported by PrS/PaS.

Although this study focussed on explaining previous findings in anterior hippocampus, it was interesting that there was little or no evidence for activation in posterior hippocampus in the imagine condition (relative to the object baseline), whereas posterior regions pSub and pCA3 were engaged for 1-week and 30-min recall respectively. Why might recalling scenes have engaged posterior hippocampus? Poppenk et al. (2013) hypothesised that posterior hippocampus represents more fine-grained or detailed spatial information than the anterior. This may be relevant as stimuli in our recall conditions were encoded visually, unlike the imagine condition, and the vividness of scenes in the 30-min recall condition was rated as higher than the other conditions. That posterior hippocampus should respond to visual stimuli as well as recalled visual stimuli extends the recent finding that visual scene perception engages posterior hippocampus significantly more strongly than imagining scenes (Zeidman et al., 2015).

These results also speak to a wider debate on the role of the hippocampus in cognition. Standard consolidation theory (Squire & Alvarez, 1995) states that the hippocampus establishes a memory trace which is then fully transferred to the cortex for long-term storage, meaning the hippocampus has no involvement in the recall of remote memories. By this account, we might expect reduced hippocampal response for 1-week recall compared to 30-min recall, but that was not the case. We found PrS/PaS was engaged by both conditions, with greater activation of uncus for 1-week recall than 30-min recall. Our results complement the findings of Bonnici et al. (2013), who found that memories over longer time periods – 2 weeks and 10 years prior to scanning – could be decoded from the subfields of anterior hippocampus. Together, these results are better supported by models which propose involvement of the hippocampus in vivid recall in perpetuity, such as multiple trace theory (Nadel and Moscovitch, 1997, Winocur and Moscovitch, 2011) and the scene construction theory (Hassabis and Maguire, 2007, Maguire and Mullally, 2013).

Imagining and recalling scenes is known to engage a network of regions known as the ‘core’ network (Addis et al., 2009, Andrews-Hanna et al., 2010, Summerfield et al., 2010, Svoboda et al., 2006). We performed an analysis of the whole (partial) volume to identify the regions which co-activated with the hippocampal subfields identified above. For the main effect of scenes, PrS/PaS was accompanied by activation of bilateral PHC, RSC, vmPFC and lateral temporal cortex. A conjunction analysis showed this wide network to be engaged in common between imagining novel scenes and recalling scenes from a week earlier, supporting our hypothesis that a common process, which we suggest is scene construction, underpins both tasks. Although we found differences in uncus and aSub between imagination and 1-week scene recall, we did not find differences between these conditions in the whole-brain results.

The response to recalling scenes from 30 min prior to scanning had similarities and differences with the other conditions. A conjunction analysis of imagining novel scenes and recalling scenes from 30 min earlier (and similarly for the conjunction of 1-week and 30 min scene recall) found that all three scene tasks engaged PHC and RSC, regions of the core network associated with spatial scene processing. As all scene tasks included the requirement to vividly simulate scenes, these results suggest an ‘inner core’ network of PrS/PaS, PHC and RSC may support this process, which speaks to previous models of PHC and RSC function in spatial imagery (Byrne et al., 2007). The main differences we identified for 30-min recall were in lateral temporal cortices. Imagining novel scenes and recalling scenes from one week earlier each engaged regions of anterior STS and MTG, and there was a significantly reduced response in these regions during 30-min recall. What might STS and MTG contribute to the process of scene construction, which we suggest explains their increased involvement during imagination and 1-week recall of scenes?

This region has wide anatomical connectivity – within the core network it is directly connected with vmPFC and PHC, whereas more posterior STS is connected with RSC (Binder, Desai, Graves, & Conant, 2009). The anterior STS in particular has been implicated in autobiographical memory (Svoboda et al., 2006) and future thinking (Schacter, Addis, & Buckner, 2007). Anterior temporal cortex is particularly associated with semantic processing, and it may act as an amodal hub for linking together information of different modalities into unified concepts (Binder et al., 2009, Patterson et al., 2007, Ralph et al., 2010). Part of the definition of a scene is that it should be coherent (Maguire & Mullally, 2013), and as such when a scene is constructed, only relevant and meaningful items should be included. We may speculate that STS and MTG could provide the semantic information required to decide which elements make sense in the context of a scene. This process could be mediated by vmPFC, as the two regions have monosynaptic connectivity and lesions to vmPFC result in confabulation (Gilboa, Moscovitch, Baddeley, Kopelman, & Wilson, 2002), where patients spontaneously generate narratives of events that never occurred.

We used fMRI to investigate the responses in hippocampal subregions and the wider brain to imagining and recalling scenes. Together, our results enable us to extend the ‘core’ network for scenes/memory to include specific subregions of the hippocampus, and propose a subdivision of this network into functional sub-networks. When imagining or recalling scenes, PrS/PaS is engaged together with PHC and RSC. These regions facilitate representation of the scenes and the production of simulated sensory imagery, as well as supporting the recall of scenes that have recently been encoded and have not yet been consolidated. However, if the scenes are recalled from consolidated memories or are newly constructed in the imagination, then vmPFC and lateral temporal cortex are also recruited, together with further subdivisions of the hippocampus (particularly aSub, pSub and uncus).

These results set clear directions for future work. First, what is it about scenes that cause certain regions within the hippocampus to be engaged more than single objects, and how exactly is this achieved? Second, it would help to know whether specific hippocampal subregions show a parametric increase in response with the age of memories being recalled. And third, what is the functional connectivity between the hippocampal subregions we have observed here and the rest of the brain? An analysis of this kind could help to better understand the role of the uncus, aSub and PrS/PaS in scene construction and scene recall.